JP2014050590A - Action recognition device for hand, artificial hand operation device, and composite sensor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve recognition accuracy of an intended action of a hand without increasing the number of wearing parts of surface electrodes each measuring a surface myoelectric potential.SOLUTION: An action recognition device 10 for a hand includes: a signal measurement part 12 measuring myoelectric potentials and muscle protuberance amounts in a plurality of different positions in the circumference of a forearm, and outputting myoelectric potential signals and muscle protuberance signals; an integration myoelectric arithmetic part 14 calculating an integration myoelectric signal on the basis of the measured myoelectric potential signal; a characteristic extraction part 16 extracting characteristic information from the myoelectric potential signal, the integration myoelectric signal and the muscle protuberance signal in each position in the circumference of the forearm; and an action identification part 22 identifying an intended action on the basis of the characteristic information of the myoelectric potential signals, the integration myoelectric signals and the muscle protuberance signals extracted by the characteristic extraction part 16.

Description

  The present invention relates to a device for recognizing an intended hand movement, a prosthetic hand operation device using the device, and a composite sensor that can be used in the device and can simultaneously measure surface myoelectric potential and muscle uplift.

  The method of estimating the intended hand movement is a core technology of the human interface that transmits the human movement intention to the machine. As a method for estimating the motion intention of the hand, there is a method for estimating the motion intention of the hand based on the surface myoelectric potential. The surface myoelectric potential is a biological signal that can be measured by an electrode attached to the skin surface, and it is possible to know which muscle is active and how much depending on the operation. The method using surface myoelectric potential uses myoelectric potential of the forearm that can acquire hand movement information, so there is no need to wear sensors that hinder hand movement on the hand itself, and there is flexibility in the usage environment . In addition, in the technique using surface myoelectric potential, if the surface myoelectric potential can be measured from the arm, it is possible to estimate the intention of the hand even when the hand is cut, which is a powerful tool for operating the electric prosthetic hand. The fact that it can be used as a control signal is a major feature not found in methods based on camera images and bending sensor information.

  A method for estimating the movement intention of the hand from the surface myoelectric potential has been studied since the 1940s. In the 1990s, many methods based on pattern recognition such as a neural network were proposed. After 2000, for example, as described in Non-Patent Documents 1 to 3, the number of motions that can be estimated and the accuracy were improved. As disclosed in Non-Patent Document 4, the present inventors also estimated the motion intention of the hand using a support vector machine (Support Vevtor machine), which is one method of pattern recognition based on the surface myoelectric potential of the forearm. We have developed a method that can estimate the seven movements of the hand with high accuracy. Moreover, it has been clarified that the above method is an effective method even for forearm amputees who have lost their hands.

Daikawa Nishikawa, Wei Wei, Hiroshi Yokoi, Yasunobu Kaji, “Online Learning Method for Controlling Myoelectric Prosthetic Hand”, IEICE Transactions D-II, vol.J82-D-II, no.9, pp .1510-1519, 1999. O.Fukuda, T. Tsuji, M. Kaneko, and A. Otsuka, "A human-assisting manipulatorteleoperated by EMG signals and arm motions", IEEE Trans. Robot. Automat., Vol.19, no.2, pp.210 -222, 2003. Y. Huang, KB Englehart, B. Hudgins, and ADC Chan, "A gaussian mixturemodel based classification scheme for myoelectric control of powered upper limbprostheses", IEEE Trans. Biomed. Eng., Vol. 52, no.11, pp.1801 -1811, 2005. Yoshikawa Masahiro, Mikawa Masahiko, Tanakayo, "Real-time hand movement recognition using support vector machine using myoelectric potential", IEICE Transactions D., vol.J92-D, no.1, pp.93- 103,2009. Masahiro Yoshikawa, Tomoki Mita, Masahiko Mikawa, Seino Tanaka, “Fundamental study on hand motion recognition based on myoelectric signal for forearm amputees”, Ergonomics, vol.46, no.3, pp .197-207,2010.

  With surface myoelectric potentials that have been used in the past, in principle, the activity of muscles existing deep in the human body cannot be observed. In addition, muscle activity related to forearm pronation or pronation that can be measured from the forearm proximal is weak, and characteristic myoelectric potential patterns may not be obtained, resulting in misidentification with other movements. Cheap. For example, the supination motion during flexion of the elbow involves the upper arm muscle more strongly than the forearm, the forearm muscle activity is weak, and misidentification with other motions is likely to occur. In addition, since the circular pronation muscle is a muscle deep in the human body and the square pronation muscle is a muscle that exists in the distal part of the forearm, the measurement of these muscle activities is performed with a surface electrode placed proximal to the forearm, Difficult to do. For this reason, conventionally, the shortage of information is compensated by arranging surface electrodes not only on the forearm but also on the upper arm or increasing the number of surface electrodes themselves. However, if an attempt is made to improve the recognition accuracy of the intended motion without increasing the number of places where the surface electrodes are attached, some information that further supplements the information obtained from a single sensor is required.

  Therefore, an object of the present invention is to solve the problems existing in the prior art and improve the recognition accuracy of the intended hand movement without increasing the number of surface electrode mounting locations for measuring surface myoelectric potential.

  In view of the above-mentioned object, the present inventor measures the amount of bulging of the muscle that occurs when the surface myoelectric potential and muscle strength are exerted, and supplements the information on the surface myoelectric potential by simultaneously using the bulging information of the muscle to recognize the intended motion. Improved.

  That is, the first aspect of the present invention is a hand motion recognition device for recognizing an intended hand motion, and measures myoelectric potential and muscle uplift at a plurality of different positions around the forearm. A signal measuring unit that outputs a myoelectric signal and a muscle bulge signal, an integrated myoelectric calculation unit that calculates an integrated myoelectric signal based on the measured myoelectric signal, and at each position around the forearm A feature extraction unit that extracts feature information from the myoelectric signal, the integrated myoelectric signal, and the muscle uplift signal, respectively, and the myoelectric signal, the integrated myoelectric signal, and the muscle uplift signal extracted by the feature extraction unit A motion recognition device for hand is provided that includes a motion identification unit that identifies an intended motion based on the feature information.

  The hand motion recognition device includes: a learning data generation unit that generates learning data from feature information acquired in advance for each of a plurality of predetermined hand movements; and learning generated by the learning data generation unit A support vector machine learning unit that configures a support vector machine based on the data, and the motion identification unit is extracted by the feature unit by data processing by a support vector machine configured by the support vector machine learning unit. It is preferable to identify an intended operation from the feature information.

  In the hand motion recognition apparatus, the signal measuring unit measures myoelectric potentials using myoelectric sensors mounted at different positions around the user's forearm, and It is preferable to measure the amount of muscle uplift using distance sensors mounted at different positions away from the skin surface of the user's forearm. In this case, it is more preferable that the signal measuring unit measures a myoelectric potential and a muscle bulge amount using a composite sensor having both the myoelectric potential sensor and the distance sensor.

  Further, the feature information includes a cepstrum coefficient obtained by Fourier transform of a myoelectric signal within a predetermined time, and an average integrated myoelectric potential obtained by averaging the integrated myoelectric signal within a predetermined time. It is preferable that the feature vector includes at least a component of an average muscle elevation obtained by averaging muscle elevation signals within a predetermined time.

  Moreover, this invention is a prosthetic hand operation apparatus for operating a prosthetic hand according to the operation | movement of the intended hand as 2nd aspect, Comprising: The said operation | movement recognition apparatus for hands, and the control apparatus which controls operation | movement of a prosthetic hand A prosthetic hand operating device in which the control device moves the prosthetic hand according to an intended motion recognized by the motion recognition device.

  Furthermore, the present invention provides, as a third aspect, a support portion, a pair of myoelectric electrodes provided to be parallel to the bottom surface of the support portion in order to detect surface myoelectric potential, and a skin surface. And a distance sensor unit supported by the support unit so as to measure the surface myoelectric potential by the myoelectric electrode and measure the muscle bulge amount by the distance sensor unit. I will provide a.

  According to the hand motion recognition device of the present invention, the prosthetic hand operation device using the same, and the composite sensor, in addition to the myoelectric potential on the skin surface, the amount of muscle uplift is measured. It is possible to obtain information on the amount of muscle uplift that occurs when muscular strength is exerted, and by simultaneously using information on the amount of muscle uplift to supplement the information on myoelectric potential, the number of attachment points of surface electrodes that measure surface myoelectric potential is not increased. In addition, the recognition accuracy of the motion intention of the hand can be improved.

It is a block diagram which shows the whole structure of the operation recognition apparatus for hands by this invention. It is a schematic diagram of a method of recognizing a motion intention by a manual motion recognition device according to the present invention. The composite sensor suitable for use as a signal measurement part of the operation | movement recognition apparatus for hands shown by FIG. 1 is shown, (a) is a fixed view, (b) is a side view. It is a graph which shows the relationship between the distance of the optical distance sensor part of the composite sensor shown by FIG. 3, a user's skin surface, and the output voltage from an optical distance sensor part. It is explanatory drawing which shows the mounting position of the composite sensor shown by FIG. It is explanatory drawing which shows an example of the mounting method of a composite sensor. It is a graph which shows the output signal obtained from a composite sensor at the time of operation | movement of various hands, (a) is a myoelectric signal, (b) is an integrated myoelectric signal, (c) is a graph regarding a sensor-skin distance signal. It is. It is a table | surface which shows the recognition rate of the motion of the hand at the time of using an average integral myoelectric feature, a cepstrum feature, and an average sensor-skin distance feature. It is a table | surface which shows the recognition rate of the motion of the hand at the time of using an average integral electromyography feature and a cepstrum feature. 4 is a graph showing a recognition result of an intention of action for a subject C. It is a block diagram which shows the whole block diagram of the artificial hand operating device using the operation | movement recognition apparatus for hands by this invention.

Hereinafter, a hand motion recognition apparatus according to the present invention will be described with reference to the drawings.
First, referring to FIG. 1, the overall configuration of a manual motion recognition device according to the present invention will be described. The hand motion recognition device 10 according to the present invention includes a signal measurement unit 12, an integral myoelectric calculation unit 14, a feature extraction unit 16, a learning data generation unit 18, a support vector machine learning unit (hereinafter referred to as "SVM learning"). Part).) 20 and an action identifying unit 22.

  The signal measuring unit 12 measures the myoelectric potential on the skin surface and the amount of muscle uplift (hereinafter referred to as “muscle uplift amount”) at a plurality of different positions around the user's forearm, and represents the myoelectric potential. A signal (hereinafter, referred to as “myoelectric signal”) and a signal indicating the muscle elevation amount information (hereinafter, referred to as “muscle elevation signal”) are output. The integrated myoelectric calculation unit 14 obtains an integrated myoelectric signal from the myoelectric signal by calculation. The feature extraction unit 16 extracts feature information from the myoelectric signal and myomy signal output from the signal measurement unit 12 and the integrated myoelectric signal output from the integration myoelectric calculation unit 14 in units of frames (within a certain time). Signal) and integrated as a feature vector. Specifically, the feature information includes a cepstrum coefficient obtained by Fourier transform of the myoelectric signal, an average integrated myoelectric potential obtained by averaging the integrated myoelectric signal, and an average obtained by averaging the myoelectric signal. The amount of muscle uplift. The learning data generation unit 18 acquires in advance a feature vector that includes the feature information of the measurement data when each of a plurality of predetermined hand movements is to be performed, and assigns an action class to these feature vectors. To generate learning data. The SVM learning unit 20 configures an identification function from the learning data generated by the learning data generation unit 18 using a support vector machine. The action discriminating unit 22 uses the myoelectric signal, the integrated myoelectric potential signal, and the feature vector of the muscle protuberance signal obtained from the measurement of the myoelectric potential and the muscle uplift amount by the signal measuring unit 12 using the discriminant function configured by learning. Identify the intended behavior (motion intention).

  As shown in FIG. 2, the motion recognition device 10 includes a myoelectric signal and a muscle uplift signal (muscle uplift information (sensor-skin distance information) obtained by a signal measuring unit 12 such as a composite sensor attached to the forearm. ) And the integrated myoelectric signal calculated from the myoelectric signal by the integrating myoelectric calculation unit 14, feature information is extracted by the feature extraction unit 16 and integrated into a feature vector. When recognizing an action, an action class to which the feature vector belongs is determined by an identification function learned in advance using a support vector machine, and the action intended by the user is recognized.

Below, each component of the action recognition apparatus 10 is demonstrated in detail.

  The measurement of the myoelectric potential and the amount of muscle uplift by the signal measuring unit 12 is performed, for example, by measuring the myoelectric potential with a myoelectric electrode brought into contact with the skin surface and using a distance sensor arranged at a predetermined distance from the skin surface It can be realized by measuring the amount of muscle uplift. In the present embodiment, the myoelectric potential and the muscle bulge amount are measured by the composite sensor 24 as shown in FIG. However, it goes without saying that the myoelectric electrode and the distance sensor may be used separately to measure the myoelectric potential and the muscle bulge amount.

  The composite sensor 24 shown in FIG. 3 is disposed away from the skin surface of the user, the support portion 24a, a pair of myoelectric electrodes 24b provided on the bottom surface of the support portion 24a so as to be parallel to each other. The distance sensor unit 24c is supported by the support unit 24a, and the myoelectric electrode 24b measures the myoelectric potential on the skin surface and outputs a myoelectric signal. By detecting the distance from the sensor unit 24 to the skin surface, the amount of muscle protrusion is measured and a muscle protrusion signal is output. The myoelectric potential electrode 24a is preferably composed of a silver electrode. The myoelectric signal measured by the myoelectric electrode 24a is filtered in a band of 10 to 482 Hz and amplified by 2000 times. The distance sensor unit 24c preferably has an infrared light emitting unit and a light receiving unit, and an optical distance sensor that measures the distance by detecting the degree of reflection of the emitted infrared light is preferably used. FIG. 4 shows the relationship between the distance between the optical distance sensor and the user's skin surface and the output voltage from the optical distance sensor when an optical distance sensor is used as the distance sensor unit 24c. Yes. It can be seen that when the distance between the optical distance sensor and the user's skin surface is 0 mm to 3 mm, the relationship between the distance and the output voltage is almost linear, and the distance up to 6 mm can be measured. In the present embodiment, the distance sensor unit 24c is supported by the support unit 24a so that the distance from the optical distance sensor (distance sensor unit 24c) to the skin surface of the subject in the neutral position is 5 mm. However, as long as the distance to the skin surface of the user is measurable, the distance between the distance sensor unit 24c and the skin surface of the subject in the neutral position can be arbitrarily set.

  In this embodiment, the signal measuring unit 12 recognizes the motion intention of seven hands using the four composite sensors 24. An example of the arrangement of the composite sensor 24 in this case is shown in FIG. In FIG. 5, the first composite sensor 24 (Ch. 1) is directly above the ulnar carpal flexor, the second composite sensor 24 (Ch. 2) is directly above the deep finger flexor, and the third composite sensor 24. (Ch.3) is disposed immediately above the heel side carpal extensor and the fourth composite sensor 24 (Ch.4) is disposed immediately above the total finger extensor. For example, as shown in FIG. 6, the composite sensor 24 may be fixed around the forearm with a band or the like. The distance sensor unit 24c of the composite sensor 24 is exposed from the band, and is arranged so that the distance sensor unit 24c and the skin surface are separated by a distance of about 5 mm in the neutral position.

  The integrated myoelectric calculation unit 14 obtains an integrated myoelectric potential signal based on the myoelectric signal obtained by the signal measuring unit 12. The integral myoelectric calculation unit 14 may obtain an integral myoelectric signal that is an envelope waveform by smoothing the myoelectric signal with a low-pass filter having full-wave rectification and a cutoff frequency of 2.4 Hz, for example. However, it is also possible to obtain the integrated myoelectric signal by other methods such as obtaining the signal with an integrating muscle by integrating the obtained myoelectric signal after full-wave rectification.

  The myoelectric signal and the muscle uplift signal obtained by the signal measuring unit 12 are sampled at a predetermined period (for example, 2000 Hz) by a 16-bit AD converter, for example. Myoelectric signal and bulge signal measured by the composite sensor 24 when palm flexion, dorsiflexion, grasping, opening, pronation, and pronation are performed, and integrated myoelectric calculation calculated from the myoelectric signal An example of the issue is shown in FIG. FIG. 7A shows the myoelectric signal, FIG. 7B shows the integrated myoelectric signal, and FIG. 7C shows the sensor-skin distance information, that is, the muscle uplift signal.

  The feature extraction unit 16, the learning data generation unit 18, the SVM learning unit 20, and the action identification unit 22 can be realized, for example, by executing a program with a CPU of a personal computer.

  The feature extraction unit 16 extracts feature information from each sampled signal in units of frames. More specifically, the feature extraction unit 16 shifts the frame at a frame length of 64 ms and a frame period of 16 ms, and uses a cepstrum feature (hereinafter referred to as “cepstrum feature”) from each of the myoelectric signal, the integrated myoelectric signal, and the muscle uplift signal cut out in units of frames. , “CC feature”), average integrated myoelectric feature (hereinafter referred to as “AIEMG feature”), sensor-skin distance feature (hereinafter referred to as “ASSD feature”). .

The AIEMG feature is an intra-frame time average of the integrated myoelectric signal, and represents the magnitude of the amplitude of the myoelectric signal. Within the p-th (p = 1,..., P) frame, the n-th integrated myoelectric signal sample is taken as IEMG _l (n) (n = 0,..., N−1; l = 1,. ). Here, N indicates the number of samples in one frame, and L indicates the number of sensors. In this embodiment, N = 128 and L = 4. AIEMG _l (p), which is the AIEMG feature obtained from the l-th sensor, is obtained by the following equation (1).

で表すと、ケプストラム係数ＣＣ_l ⁿ（ｐ）は次式（３）により求められる。

The CC feature is a cepstrum coefficient obtained by cepstrum analysis, and is obtained as follows by cepstrum analysis with respect to the myoelectric signal in the frame. EMG signal samples n-th point in the p-th frame as the EMG _l (n), _EMG l Fourier transform ^{_{X l k (p) (k}} = 0, ..., N-1) of the (n) the

, The cepstrum coefficient CC _l ⁿ (p) is obtained by the following equation (3).

  The cepstrum analysis can separate the envelope shape and the fine structure of the power spectrum, and the characteristic of the envelope shape appears in the lower order and the feature of the fine structure appears in the higher order coefficient. The feature extraction unit 16 employs W low-order coefficients from n = 0 to n = W−1 as CC features, and uses an envelope shape as an extraction feature. Since the cepstrum coefficient is calculated using a fast Fourier transform, it is suitable for real-time calculation. In this embodiment, W = 3 is adopted as an example.

ここで、ＳＳＤｌ（ｎ）（ｎ＝０，…，Ｎ−１；ｌ＝１，…，Ｌ）は、l番目のセンサから得た、第ｐ（ｐ＝１，…，Ｐ）番フレーム内のｎ点目の筋隆起信号サンプルである。 The ASSD feature ASSD _l (p) is obtained by an intra-frame time average of a muscle bulge signal that represents the distance between the sensor and the skin surface, as shown in Equation (4) below.

Here, SSDl (n) (n = 0,..., N-1; l = 1,..., L) is in the pth (p = 1,..., P) th frame obtained from the lth sensor. This is an nth muscle uplift signal sample.

The AIEMG feature, CC feature, and ASSD feature thus obtained are integrated into the following feature vector x (p).

  In the present embodiment, the one-dimensional AIEMG feature, the three-dimensional CC feature, and the one-dimensional ASSD feature are extracted from each channel (that is, each composite sensor 24) and integrated as a feature vector, so that a 20-dimensional feature vector is obtained. . The feature vector is used after smoothing by moving average using past frame data. For example, the feature vector is smoothed by a moving average of 11 points using frame data from the past 10 points to the present.

  The learning data generation unit 18 acquires in advance feature vectors extracted from measurement signals when each of a plurality of predetermined hand movements is to be performed, and assigns action classes to these feature vectors, Generate learning data. The method of assigning a motion class based on motion information separately measured by an indirect angle sensor or the like is difficult to apply to a forearm amputee who cannot directly measure hand movement. Therefore, the learning data generation unit 18 of the motion recognition apparatus 10 according to the present embodiment assigns the motion class using only the measurement data of the myoelectric potential when the motion is performed in a predetermined order. For this reason, it is necessary to detect a time interval (hereinafter referred to as “motion interval”) in which the muscle contraction is made with the intention of motion.

Below, an example of the production | generation method of learning data is described. However, the learning data generation method is not limited to the following.
First, the myoelectric potential measurement data, that is, the myoelectric signal obtained continuously in a predetermined order with the neutral position between the motions to be recognized is obtained, and the feature vector x (p) is obtained. In each operation, the joint is moved from the neutral position to the automatic joint range of motion, and then the force is released to return to the neutral position again. Next, the motion section is detected based on the zero crossing number of the myoelectric signal. Based on the definition of zero crossing shown in the following equation (6), the EMG _l ^S in-frame zero crossing number ZC _l (p) obtained by normalizing the myoelectric signal measured by the myoelectric potential electrode 24b with the maximum value. Is obtained by calculation.

ここでは、Ｍ＝１０とした。１回の筋電信号データ内のＺＣ_ma（ｐ）には動作を行った回数だけ極大点が存在するので（中立位は除く）、各極大点のＺＣ_ma（ｐ）に０.５をかけた値以上の区間を動作区間として、それぞれの動作区間内の特徴ベクトルに対して同一の動作クラスを付与する。 Next, a sum ZC _sum (p) of ZC _l (p) is obtained and further smoothed by a moving average.

Here, M = 10. Since ZC _ma (p) in one EMG signal data has local maximum points as many times as the number of movements (excluding neutral position), 0.5 is applied to ZC _ma (p) at each local maximum point. The same motion class is assigned to the feature vector in each motion section, with a section greater than the specified value as the motion section.

ここで、ｙ_iはｉ番目の学習サンプルｘ_iに対応するクラスラベルであり、λ_iはラグランジュの未定乗数、ｂはバイアス項である。また、Ｋ（ｘ_i、ｘ）はカーネル関数である。 The SVM learning unit 20 configures an identification function from the learning data generated by the learning data generation unit 18 using a support vector machine. The support vector machine learns parameters from the learning data on the basis of “margin maximization”. An identification function that identifies one of two classes of an unknown pattern feature vector x (p) (hereinafter simply referred to as x) is expressed by the following equation (8).

Here, y _i is a class label corresponding to the i-th learning sample x _i , λ _i is a Lagrange undetermined multiplier, and b is a bias term. K (x _i , x) is a kernel function.

ここで、γはカーネルパラメータである。カーネル関数は線型分離不可能な学習データを高次元の特徴空間に写像し、写像先の特徴空間において線型分離可能にする。そのため、線型分離不可能な分布になりやすい筋電位パターンの識別に適している。 In the SVM learning unit 20 of the motion recognition apparatus 10 of this embodiment, one kernel parameter to be determined is easily obtained, and the discrimination performance is better than that of the sigmoid kernel and polynomial kernel. The represented RBF (radial basis function) kernel is used. However, other kernel functions can be used in the SVM learning unit 20.

Here, γ is a kernel parameter. The kernel function maps learning data that cannot be linearly separated into a high-dimensional feature space, and enables linear separation in the feature space of the mapping destination. Therefore, it is suitable for identifying myoelectric potential patterns that tend to have a distribution that cannot be linearly separated.

上記の式（１０）は凸二次計画問題であり、大局的最適解が保証される。求めたλ_iのうち、非０のλに対する学習サンプルはサポートベクターと呼ばれ、識別関数は学習サンプル中の少数のサポートベクターのみで構成されるため、識別に必要な計算量は少なくなる。Ｃはどの程度の誤識別を許すかを決定するペナルティーパラメータであり、学習時に予め定めておくハイパーパラメータは、このＣとカーネルパラメータγの二つのみである。これにより、ハイパーパラメータの探索が容易となっている。 In order to actually obtain the discriminant function shown in Equation (8), λ _i that maximizes the following equation is obtained according to the margin maximization criterion.

The above equation (10) is a convex quadratic programming problem, and a global optimum solution is guaranteed. Of the obtained λ _i, the learning sample for non-zero λ is called a support vector, and the discriminant function is composed of only a small number of support vectors in the learning sample, so the amount of calculation required for discrimination is reduced. C is a penalty parameter that determines how much misidentification is allowed, and there are only two hyperparameters C and kernel parameter γ that are determined in advance during learning. This facilitates searching for hyperparameters.

  Since the support vector machine is a technique for discriminating two classes in principle, a “one-against-one” algorithm is used to identify multiple classes. In this algorithm, all combinations of O classes, that is, O (O + 1) / 2 discriminant functions are constructed, and feature vectors are discriminated using each discriminant function. All the identification results are aggregated, and the action class identified most is the final identification result in the frame. The SVM learning unit 20 may use the “one-against-all” algorithm instead of the “one-against-one” algorithm as a method for extending the support vector machine to multiple classes.

[Motion recognition experiment]
In order to confirm the effectiveness of the motion recognition device 10, a motion recognition experiment was performed. The test subjects were 3 men and 1 woman for a total of 4 people, and the procedure was as follows. The subject sits in a natural posture on a chair placed in front of the computer display. The elbow was separated from the body by about 15 cm and bent about 80 degrees so that the elbow and the hand were in the same horizontal position. Also, the forearm and hand were kept out of contact with the desk. One trial was 60 seconds, and a total of 30 operations were performed in the order of palm flexion, dorsiflexion, grasping, opening, forearm pronation, and forearm prolapse. In each operation, after moving the joint from the neutral position to the range of automatic joint movement, the force was removed and the neutral position was returned to the neutral position, and the next operation was performed. It took about 1 second from the start to the completion of each operation. In addition to the myoelectric signal and the integrated myoelectric signal, when using the myoelectric signal (that is, the distance information between the sensor and the skin surface) and when using only the myoelectric signal and the integrating myoelectric signal without using the myoelectric signal ( In other words, a total of 10 trials were performed for each of the cases where only information from the myoelectric sensor was used. Then, in each of them, a motion recognition experiment was performed using 2 trials out of 10 trials as learning data and the remaining 8 trials as test data, and the results of both were compared. For the evaluation, the motion recognition rate obtained below was used.

  The motion recognition rate for each subject performed in this way is shown in FIGS. FIG. 8 shows the result of the motion recognition rate when using feature information extracted from the myoelectric signal, the integrated myoelectric signal, and the muscle uplift signal, that is, a feature vector that combines the CC feature, the AIEMG feature, and the ASSD feature. FIG. 9 shows a case where a feature vector combining only CC features and AIEMG features extracted from the myoelectric signal and the integrated myoelectric potential signal is used except for the muscle uplift signal (that is, information on the sensor-skin surface distance). This is a result of the motion recognition rate. When a feature vector combining the CC feature, the AIEMG feature, and the ASSD feature was used, the total recognition rate of 7 actions was 95% or more for all the subjects. In addition, compared with the case where the feature vector combining only the CC feature and the AIEMG feature is used, the total recognition rate when using the feature vector combining the CC feature, the AIEMG feature and the ASSD feature is improved by about 1%. there were. Although the overall movement was improved by about 1%, the recognition rate was particularly improved when grasping, opening, forearm gyrus, and forearm gyrus. From this result, the information about the myoelectricity, i.e., the myoelectric signal and the integral myoelectric signal, is added to the information about the amount of bulging of the muscle (distance between the sensor and the skin surface), i.e., the muscle bulging signal, and only the information about myoelectricity is obtained. It can be seen that motion intention related information that can be supplemented.

  FIG. 10 shows an example in which motion recognition is performed on the subject C in the motion recognition experiment. The subject repeatedly performed palm flexion, dorsiflexion, grasping, opening, forearm pronation and forearm prolapse. FIG. 10 shows that the intended motion is recognized with high accuracy.

  In this way, since the motion recognition device 10 measures the amount of muscle protrusion from the signal measuring unit 12 such as the composite sensor 24 in addition to the myoelectric potential on the skin surface, in addition to the myoelectric signal and the integrated myoelectric signal, Based on the feature vector extracted by the feature extraction unit 16 from the muscle uplift signal, the operation identification unit 22 identifies the intended operation using the identification function learned by the SVM learning unit 20. . That is, it is possible to supplement the information obtained from the myoelectric signal and the integrated signal by simultaneously using the muscle bulge signal, and to improve the recognition accuracy of the intended hand movement without increasing the number of measurement points by a sensor or the like. it can.

  Next, the prosthetic hand operating device 30 using the motion recognition device 10 shown in FIG. 1 will be described. The prosthetic hand operating device 30 includes the motion recognition device 10 including the composite sensor 24, a control unit 32, and a prosthetic hand driving unit 34. The composite sensor 24 of the motion recognition device 10 is attached to the forearm, measures the myoelectric potential on the skin surface of the user's forearm and the amount of muscle uplift, and outputs a myoelectric signal and a muscle uplift signal. The motion recognition apparatus 10 calculates an integrated myoelectric signal based on the myoelectric signal output from the composite sensor 24, and includes a CC feature, an AIEMG feature, and an ASSD feature from the myoelectric signal, the integrated myoelectric signal, and the myoelectric signal. Feature information is extracted, and an intended operation is recognized using a support vector machine based on learning data obtained in advance and feature information obtained from measurement data of the composite sensor 24. The control unit 32 controls the prosthetic hand driving unit 34 so as to operate the prosthetic hand according to the motion recognized by the motion recognition device 10.

  The prosthetic hand operating device 30 configured as described above uses the motion recognition device 10 and can accurately recognize the motion intended by the user. It becomes possible to make it.

  As described above, the motion recognition device 10 and the prosthetic hand operating device 30 using the same according to the present invention have been described based on the illustrated embodiment, but the configurations of the motion recognizing device 10 and the prosthetic hand operating device 30 of the present invention are illustrated. However, the present invention is not limited to the embodiment. For example, in the illustrated embodiment, the composite sensor 24 is used as the signal measuring unit 12 of the motion recognition device 10. However, instead of the composite sensor 24, myoelectric communication is performed using individual myoelectric potential sensors and distance sensors. No. and muscle uplift signal may be acquired.

DESCRIPTION OF SYMBOLS 10 Action recognition apparatus 12 Signal measurement part 14 Integral myoelectric calculation part 16 Feature extraction part 18 Learning data generation part 20 SVM learning part 22 Action identification part 24 Composite sensor 24a Support part 24b Myoelectric electrode 24c Distance sensor part 30 Prosthetic hand operating device 32 Control unit 28 Prosthetic hand drive unit

Claims (7)

A hand motion recognition device for recognizing an intended hand motion,
A signal measuring unit that measures myoelectric potential and muscle bulge amount at a plurality of different positions around the forearm, and outputs a myoelectric signal and a muscle bulge signal;
An integrated myoelectric calculator that calculates an integrated myoelectric signal based on the measured myoelectric signal,
A feature extraction unit that extracts feature information from each of the myoelectric signal, the integrated myoelectric signal, and the muscle bulge signal at each position around the forearm;
An action identifying unit that identifies an intended action based on feature information of the myoelectric signal, the integrated myoelectric signal, and the muscle protuberance signal extracted by the feature extracting unit;
A hand motion recognition device comprising:   A learning data generation unit that generates learning data from feature information acquired in advance for each of a plurality of predetermined hand movements, and a support vector machine is configured based on the learning data generated by the learning data generation unit A support vector machine learning unit that performs the intended operation from the feature information extracted by the feature extraction unit by data processing by a support vector machine configured by the support vector machine learning unit. The hand motion recognition device according to claim 1, wherein   The signal measuring unit measures myoelectric potentials using myoelectric potential sensors mounted at different positions around the user's forearm, and from the skin surface of the user's forearm to different positions around the user's forearm. The motion recognition device for a hand according to claim 1 or 2, wherein the amount of bulging muscle is measured using a distance sensor mounted apart.   The hand motion recognition device according to claim 3, wherein the signal measurement unit measures a myoelectric potential and a muscle bulge amount using a composite sensor that combines the myoelectric potential sensor and the distance sensor.   The feature information includes a cepstrum coefficient obtained by Fourier transform of a myoelectric signal within a predetermined time, an average integrated myoelectric potential obtained by averaging integrated myoelectric signals within a predetermined time, The manual action according to any one of claims 1 to 4, wherein the motion vector is a feature vector including at least a component of an average muscle elevation obtained by averaging muscle elevation signals within a predetermined time. Recognition device. A prosthetic hand operating device for operating a prosthetic hand according to the intended hand movement,
The hand motion recognition device according to any one of claims 1 to 5,
A control device for controlling the operation of the prosthetic hand;
The prosthetic hand operating device is characterized in that the control device moves the prosthetic hand according to the intended motion recognized by the motion recognition device.   A support unit, a pair of myoelectric electrodes provided parallel to each other on the bottom surface of the support unit to detect surface myoelectric potential, and a support unit supported by the support unit so as to be spaced apart from the skin surface And a distance sensor unit that measures a surface myoelectric potential by the myoelectric electrode and measures a muscle bulge amount by the distance sensor unit.

Images